Computation-Limited Bayesian Updating: A Resource-Rational Analysis of Approximate Bayesian Inference

Jian-Qiao Zhu, Thomas L. Griffiths

TL;DR

This paper proposes that human "conservatism"—the tendency to underweight new evidence when updating beliefs—is actually a mathematically optimal strategy when computational resources are limited. By formalizing the "cost of thinking" using information theory, the authors derive a new belief-updating rule that perfectly mirrors observed human biases and bridges the gap between abstract Bayesian theory and real-world cognitive constraints.

What Problem Does This Paper Address?

In a perfect world, we would all be "Ideal Bayesian Agents," updating our beliefs using Bayes' rule to perfectly integrate new data with our prior knowledge. However, decades of psychological research show that humans are "conservative": we don't change our minds as much as Bayes' rule says we should. When shown a lopsided sample of red chips from an urn, we might estimate a 75% probability of it being the "mostly red" urn, even when the math dictates 97%.

Previously, researchers explained this using "heuristics and biases"—essentially saying our brains are buggy. More recently, "rational process models" suggested we use clever but noisy approximations like sampling. This paper addresses the fundamental why: Is there a principled, "resource-rational" reason for this systematic underreaction to data? The authors argue that because changing a belief state requires computational effort, an intelligent agent must balance the accuracy of an update against the "cost" of making that change.

Methods

The authors introduce a new objective function for belief updating grounded in Information Theory.

  1. The Objective: They model an agent who wants to minimize two things:
  2. Accuracy Error: The Kullback-Leibler (KL) divergence between their new belief and the true Bayesian posterior.
  3. Computational Cost: The KL divergence between their new belief and their old belief (representing the effort required to move the "mental needle").
  4. The Parameter $\lambda$: They introduce a conversion factor, $\lambda$, which represents the cost of processing one unit of information.
  5. The Derivation: By solving this optimization problem, they derive the Computation-Limited Bayesian Updating rule.
  6. Validation: They test this rule against two pillars:
  7. Empirical Meta-analysis: Re-analyzing over a dozen classic psychology studies on sequential belief updating.
  8. Algorithmic Simulation: Running Variational Inference (VI) simulations using a neural network (MLP) to see if limited gradient descent steps produce the same patterns predicted by their theory.

Datasets

The paper does not introduce new experimental data but relies on a massive curation of existing behavioral data:

  • Benjamin (2019) Meta-analysis: A dataset aggregating results from 8-12 major papers (e.g., Grether 1992, Holt & Smith 2009) involving "bookbag-and-poker-chip" or "ball-and-urn" tasks.
  • Developmental Data: Behavioral data from studies by Girotto & Gonzalez (2008) and L. Zhu & Gigerenzer (2006), which track how belief-updating behavior changes from preschool age through adulthood.

Key Insights and Ideas

  • The "Shrunken" Likelihood: The most striking "aha" moment is the derived update rule: $q^* \propto L^{1/(1+\lambda)} q$. Mathematically, the prior ($q$) is incorporated fully, but the likelihood ($L$)—the new evidence—is raised to a power less than 1. This provides a first-principles explanation for why we underweight evidence: it’s the optimal way to save "compute."
  • Prior vs. Likelihood Asymmetry: The theory predicts that computational limits specifically suppress the influence of new data ($\alpha < 1$) while leaving the weight on the prior ($\beta$) near 1. This matches human data far better than previous models that suggested we underweight both.
  • Natural Gradient Connection: The authors show that their resource-rational update is mathematically equivalent to a Natural Gradient step in parameter space. This links high-level cognitive theory directly to efficient optimization techniques used in machine learning.
  • Efficiency Increases with Age: The model explains developmental trends. As children grow, their "conversion factor" $\lambda$ decreases (they become more computationally efficient), which naturally leads them to look more like "Ideal Bayesians."

Key Results

  • Human Data Fit: In the "Grether regression" (a standard model for measuring belief updates), the theory predicts a likelihood weight ($\alpha$) of $1/(1+\lambda)$ and a prior weight ($\beta$) of 1. Meta-analysis of human data showed $\alpha \approx 0.33–0.38$ (significant underreaction) and $\beta \approx 0.95$ (near-perfect prior retention), strongly supporting the model.
  • Simulation Success: In neural network simulations of Variational Inference, the authors found that as they restricted the "compute" (fewer gradient steps or smaller step sizes), the network’s behavior shifted exactly as the theory predicted: the weight on the likelihood dropped, while the weight on the prior remained stable.
  • Simultaneous vs. Sequential: The model successfully explains why "underreaction" becomes more extreme as the "informational load" (sample size or data diagnosticity) increases.

Limitations and Future Work

  • Base-Rate Neglect: The current model primarily explains underreaction to data. It does not yet account for "base-rate neglect" (where people ignore the prior), which would require a weight on the prior ($\beta$) of less than 1.
  • Algorithmic Specifics: The analysis is at the "computational level." It abstracts away the specific dynamics of certain algorithms, such as the autocorrelation found in Markov Chain Monte Carlo (MCMC) sampling.
  • Alternative Cost Metrics: The authors acknowledge that KL divergence might not be the only way to measure mental effort; future work could explore other information-theoretic measures like Rényi entropy.
  • Amortized Inference: The paper suggests that "amortized inference" (learning to map data to beliefs over time) might explain other biases, a promising direction for future integration.

Summary updated

Abstract

Data and computational capacity are essential resources for any intelligent system that update its beliefs by integrating new information. However, both data and computational resources are inherently limited. Here, we introduce a new resource-rational analysis of belief updating that formalizes these constraints using information-theoretic principles. Our analysis reveals an interaction between data and computational limitations: when computational resources are scarce, agents may struggle to fully incorporate new data. The resource-rational belief updating rule we derive provides a novel explanation for conservative Bayesian updating, where individuals tend to underweight the likelihood of new evidence. Our theory also generates predictions consistent with several process models, particularly those based on approximate Bayesian inference.
BibTeX
@misc{zhu2024computation,
  author = {Jian-Qiao Zhu and Thomas L. Griffiths},
  title = {{Computation-Limited Bayesian Updating: A Resource-Rational Analysis of Approximate Bayesian Inference}},
  year = {2024},
  doi = {10.31234/osf.io/6aw3f},
  howpublished = {PsyArXiv},
}
Figure